How AI image generators are changing art and creativity

Keith Shaw: Continued advances in AI image creation tools have sparked a bit of a firestorm and backlash, with artists and big tech companies arguing over what’s right and wrong in this space.

Meanwhile, enthusiasm from end users about these new capabilities could end up costing AI companies more money than they expected. We're going to check in to see how creative the technology has gotten on this episode of Today in Tech.

Keith Shaw: Michael Todasco, I had you on the show about eight months ago when we were discussing the world of AI creativity—whether it was just about going beyond the magic tricks of image generation and similar tools.

In the last eight months, I’ve noticed that AI has gotten a lot better in the image creation space. You’ve been doing a bunch of creative experiments to gauge what’s going on. So, from your perspective—have you seen this improvement as well?

Michael Todasco: Yeah, let me give you a very specific example, Keith. Earlier this month—so, April—we’re recording this in April—I had two classes I was teaching: one on a Monday, one on a Wednesday.

On Monday, I gave a presentation covering image generation, its downsides, what it can and can’t do—all of that. Then on Tuesday, OpenAI announced ChatGPT-4o with image generation. By Wednesday’s class, I had to completely update my presentation. The world had literally changed in 48 hours.

Two sections of the same class got very different versions. That’s a real-life example of how fast this is moving. The new image generation tool is amazing, and there's so much other incredible stuff out there as well. The pace is just... wild.

If you're in the early cohort of a class, you might end up missing something that the later cohort experiences.

Keith Shaw: Yeah, and that model update really took off online. We saw people generating images of themselves in the style of Studio Ghibli—that was a big meme for a while, which brought about some issues we’ll get into.

But we also started seeing people do what I’d call “action figure” images. Have you seen that trend? It's like, “Draw yourself as an action figure,” and it does so based on your previous interactions with ChatGPT. Because of the memory features, it already knows who you are.

I tried it—and I didn’t like how it looked, so I never posted the action figure of me. I’ve got to work on what it thinks I look like.

Michael Todasco: Well, that was actually one thing I did—not as an action figure. I just said, “Hey, ChatGPT...” So for folks who don’t know, memory—or infinite memory, I think they’re calling it—was another feature OpenAI announced. It means any chat you've had with it is now remembered.

So I went in and asked, not as an action figure, but just, “What do you think I look like?” and “What do you think my family looks like?”

It was really interesting to see the images it generated. I became a generic white guy with a beard and glasses in his 40s.

I posted that on LinkedIn and said, “Hey, other white guys in their 40s—what are you getting?” And sure enough, many of them were getting results that looked a lot like me. So clearly, there’s an archetype built into the system—certain facial structures and all that.

It was relatively close, but not exactly me. I would’ve been shocked if it was, because I don’t know how it could comprehend that.

Keith Shaw: Yeah, I think it knows what I look like because I’ve uploaded pictures of myself before—saying things like, “Draw me as a podcast host,” or, “Draw me flying a plane.” So when I did the action figure, it probably just took the photo—me from the chest up—and filled in the blanks.

That’s what upset me about the result. It kind of told me I needed to lose more weight. Michael Todasco: Right?

All the things AI gets wrong. I’s now become your judgmental parent.

Keith Shaw: This is kind of a tangent, but remember when Wii Fit came out on the Nintendo Wii? There was definitely a cultural clash, Japanese vs. American sensibilities.

You’d stand on that little scale, and it would take your picture and basically say, “Yeah, you’re obese.” It had no qualms about it, no sugar-coating. Maybe AI is doing that now, too, drawing an image based on what it thinks you look like and not holding back.

Michael Todasco: That wouldn’t surprise me. A Japanese product being that direct? Yeah.

Keith Shaw: So, getting back to this new model with the Studio Ghibli stuff—it caused some issues. First, the animator himself, Miyazaki, got really upset about it for obvious reasons.

But on the other hand, users loved the creations so much that OpenAI's servers started getting overwhelmed. I think Sam Altman even had to come out and say they were experiencing delays because so many people were generating these images. What’s your take on that controversy?

Michael Todasco: It’s brilliant marketing on their part. I don’t think they went in expecting the Studio Ghibli thing to catch fire, but when it did—Sam Altman changed his profile picture to one of those images. They knew what they were doing.

I don’t know why it was Studio Ghibli, though. It could’ve been anything—Pixar, Disney princesses, whatever. There are probably dozens of styles the model could handle, especially early on. I think they’ve clamped down on things since then, and we could get into that too.

But you never know what will take off on the internet. People have been able to generate Studio Ghibli–style images in MidJourney and elsewhere for a while. But it couldn’t generate you as a Ghibli-style image quite like this new tool could.

That’s what really changed. It wasn’t just “generate a Ghibli image”; it was “make me look like a Ghibli character”—and it did a really good job at that. It could’ve been The Simpsons or anything else, but Ghibli won out.

I actually wrote about this: Studio Ghibli’s profits are only about $20 million—not bad, but relatively small. They're not a huge studio by any stretch.

So to see OpenAI gain all this value off a relatively small studio really makes you think. We need clear copyright laws in the U.S. I will say, though—in Japan, to the best of my knowledge, that’s all totally legal.

Keith Shaw: So maybe that’s why they chose that style? Because it’s legally easier to get away with? Michael Todasco: Maybe.

Again, I’m not a copyright attorney, but Studio Ghibli would have a trademark in the U.S., which is different than in Japan. It all depends on jurisdiction.

In Japan, around 2019, they basically said you can train an AI model on anything that's on the open internet and you don’t need to compensate the copyright holders. They're one of the most open countries when it comes to AI training data.

Keith Shaw: It's probably why they went with that instead of trying something like Disney, which has an army of lawyers.

Michael Todasco: Interesting fact about Disney—I wrote a piece about this recently. About a year ago, I found a list of the 40 most famous cartoon characters in America. I went into DALL·E 3, ChatGPT’s image generator at the time, and went through them one by one.

I said things like, “Can you generate a picture of Donald Duck? Betty Boop? Fred Flintstone?” And I noted the responses. Did it say no? Did it generate a workaround image that looked exactly like the character, just without naming it?

For example, I’d ask for Donald Duck, and it would say, “I can't generate that,” but then give me a duck wearing a sailor suit—it was Donald Duck without calling it that.

I repeated the experiment with the new image generation model just last week. This time, not a single Disney-licensed image was returned—not even the workaround. So clearly, they’ve put the clamps down on Disney’s IP.

Some other IP holders? Not so much. The model seemed more lenient overall than it was a year ago. So they’re clearly stricter with Disney—they know to avoid the mouse.

Keith Shaw: Obviously other models can still get away with more. I've seen people use Grok—that’s the Elon Musk one—and I don’t think they have guardrails at all.

Michael Todasco: Right, though I think it’s paid-only. IP guardrails there seem pretty light. MidJourney also feels pretty loose when it comes to that.

If you try something like this in Gemini—or whatever they’re calling it now—you’re not going to get anything. I think Google clamped down the most.

I tried one of the Chinese video models about six months ago. I can’t even remember the name. I asked it to generate “Mickey Mouse with a machine gun.” And it did it. No hesitation.

So that tells you where they’re at. They really don’t care. Even Grok—I don’t know if it would allow that. I haven’t tested it. I probably should.

I might go try Mickey Mouse with a machine gun on Grok. I assume it’ll stop me eventually, but who knows?

Keith Shaw: I want to show you some of the images—this is another reason I wanted to talk to you—just to show how good this stuff has gotten.

One of my go-to prompts is trying to get the system to draw a crossover in my mind: ALF, the puppet from the '80s, visiting The Golden Girls.

So I typed in, “Remember that classic episode?” Of course, it doesn’t exist. But here’s what it gave me. The first one was just horrible. The eyes were wrong, and ALF looked like he was six feet tall. That might’ve been Sophia in the background.

So that was from 2023. Then I tried another one—ALF came out looking like a teddy bear. The white-haired woman sort of looked like Dorothy.

Then I did one when Firefly first came out—apparently, it picked Rose—but ALF looked like a nightmare puppet from a horror movie. Like Chucky.

Okay, now prepare to have your socks blown off. Here's the one I did today. (holds up image)

That’s Dorothy.

That’s ALF. And they’re sharing cheesecake in the kitchen. It’s... yeah, the cheesecake.

Michael Todasco: The only minor complaint I have is that the cheesecake is backwards.

Keith Shaw: Maybe some people eat cheesecake backwards. You never know.

Michael Todasco: No way. People who think that’s the proper way to eat cheesecake—post in the comments and be prepared for some hate.

Keith Shaw: Oh no—my director, Chris, just waved at me. He says that’s how he eats cheesecake. Michael Todasco: What?

Keith Shaw: Yeah.

Crust first. Michael Todasco: No!

Keith Shaw: He says we’re missing out. We should try it that way. You’ve done things like this before—I remember one of your prompts to AI was about the proper way to eat a burrito.

Michael Todasco: Okay, Keith, I’m inspired. After this, I’m going to go into all the image generation models and ask them to show people eating cheesecake. I want to see what percentage are eating it backwards vs. forwards.

I’ll also go into Claude and other models and ask, “What’s the right way to eat cheesecake?” I need to find out how many say crust-first. I think Chris might be an AI.

Keith Shaw: There have been questions about that.

Michael Todasco: There’s no right or wrong way to eat something. You should know that as a human.

Keith Shaw: He says it’s like a new way of eating a Chipotle bowl.

Michael Todasco: Am I eating that wrong too? What am I missing?

Keith Shaw: I don’t know. He’s signaling something... Oh, he said you’re supposed to flip the bowl upside down.

When they give you the bowl, the aluminum top is on the top. You’re supposed to flip it over.

Michael Todasco: I’ve never seen that. Maybe check that with your models too.

Keith Shaw: We’ve gotten way off track, so let me bring us back.

One of the projects that fascinated me was your experiment: Can AI write a comic book? The first couple of attempts weren’t great, but with this latest model, it seems to be understanding more—especially the use of words and letters.

What takeaways did you have with the latest model in your comic book test?

Michael Todasco: Just to put it in perspective—the first time I ran this experiment was in 2022, before ChatGPT even existed. I used GPT-3 to write the comic book script—it was able to do that.

So I took that six-page script and used MidJourney to generate panel images. I had to prompt each panel individually, pick the best ones, go into Comic Life to lay them out, and do a lot of manual work.

There was a lot of human judgment required, and on top of that, there was no character consistency. One panel the character was skinny and gray, then in another, he was plump and black—it was all over the place.

And the story wasn’t great either. That was 2022. I tried again a few times over the past year, each time with marginal improvements.

Now, with ChatGPT-4o and the new image generation tool, it’s a totally different rendering method. If you use MidJourney, you see a diffusion process—the image starts as noise and gradually becomes clear.

ChatGPT’s model is different. It’s almost like a 3D printer or laser printer: top-down rendering. You see the image clarify from top to bottom, rather than all at once.

That process allows it to actually render words and letters legibly now—which is huge for comics.

So now I can say: “ChatGPT, I want a four-page comic book. Write the script, and then generate page one, two, three, and four.” It does that.

And it gets the word bubbles right. It gets the layout. I don’t have to design the pages anymore.

The only problem is character consistency between pages. Within a single page? It’s fine. But between pages—even within the same session—it changes.

But they’ve figured it out within one image. Believe me—within three months, they’ll figure it out across images. You’ll be able to generate 22 pages of a comic with consistent characters and story.

Keith Shaw: Have you tried rewriting the story?

Michael Todasco: I intentionally don’t. The reality is—it’s not a great writer yet. It’s adequate. Like a good high school–level writer. And I don’t mean that as an insult.

But when people buy a book or a comic, they’re expecting something more polished. This is okay for a first draft. It’s a decent framework you can improve upon.

So yes—it can write a story, but it’s not good enough to replace something you’d get from, say, Marvel or a professional comic book writer.

You're not going to say, “This is so good, I don’t need to read Fantastic Four anymore.” No one’s going to do that. Not yet.

Keith Shaw: It still feels like, with a lot of the creative stuff, it leans heavily on tropes.

And that’s because it’s trained on so much content that already exists—books, TV, movies, comics. So you start noticing that everything feels a bit... familiar.

When I did a D&D backstory, for example, it felt like every other generic character I’ve seen for that class.

Michael Todasco: But that said, Keith, here’s what I think we should do. There was an ALF comic book back in the day. But to my knowledge, there’s never been a Golden Girls comic.

So once this tech is capable, we should generate an ALF comic with Golden Girls as guest stars. We’ll evaluate it and maybe even make it a third appearance on your show.

Keith Shaw: That’s my weekend project. I’ll storyboard it or plot it out. I’ll prompt it with something like, “Write an episode of Golden Girls where ALF visits and causes a conflict.” Let’s see what it does.

Michael Todasco: Was this with the new model or the old one?

Keith Shaw: This was with the old model and the football helmets.

I haven’t tried it lately with the new one, but even in your comic example—which used the new model—the text was good but still slightly off.

Maybe we should just call AI slightly off.

Michael Todasco: Yeah, it would do things like spell one of 40 words on a page incorrectly.

Like, no one misspells what as hwat—but it would.

If you tell it to rerun, it can’t just correct that one word.

They don’t have spot editing in ChatGPT yet. MidJourney has it.

Some other tools do, too.

But I don’t think ChatGPT’s new image model has that yet.

Keith Shaw: The old one tried it—and it didn’t work very well.

Michael Todasco: Yeah, that’s probably why it’s not included right now.

But prompt adherence is improving.

I’ve started making presentation slides using image generation.

For example, I’m presenting in San Diego next week—and I’m using ChatGPT to create most of the visuals.

I’ll ask it for an Art Deco–style slide with a big “Agenda” header and five bullet points—it does that really well.

Michael Todasco: Think about that from a creativity standpoint. You can customize images for every slide.

In another presentation—an AI class for accountants—I thought of those old-school green visors they used to wear.

So I generated a robot wearing a green visor and made it the mascot for the deck.

Every slide had this robot in a different pose—presenting bullet points, sitting at a desk, whatever.

I uploaded the image each time and instructed the model to reuse it in new contexts.

That’s the level of personalization image generation is starting to unlock.

Keith Shaw: Yeah, everyone was focused on Studio Ghibli, but this tech has much more potential.

Some of the other new features that came out with the update haven’t even been explored yet.

Things like memory, better word rendering, and customization—I think more people should play with those.

Should there be any concern from the other side—beyond IP issues?

Michael Todasco: We need to figure out what copyright laws we want in the U.S.—Japan is very open, Europe less so.

As for concerns: if you’re a designer, you should be using this. You can be more efficient, more productive.

But if your whole job is to make PowerPoint templates, that role may not exist in five years.

Keith Shaw: I had to make a pitch recently, and I’m like a PowerPoint 0.5 on a 1–100 scale.

So I had ChatGPT write the pitch, generate the script, and design the slides.

I wrote the outline in Word, fed it to ChatGPT, and it did the rest. There was still some manual tweaking.

But as a novice, I loved it. If that were my full-time job though—I’d be a little nervous.

Michael Todasco: Whatever your job is, you should be using these tools. Understand what they can do, how they can help you and your clients.

If you broke your day into 15-minute chunks—email, meetings, decks—see which parts AI can help with.

If 80% of your day can be done better or faster with AI, that’s something to think about.

We’re not there yet for most jobs—but it’s coming. Then the question becomes: where do humans still add value?

When students ask what job they should prepare for, I always say: solopreneur.

One human, a bunch of AI tools, building something real for human customers—and being the human in the loop when needed. Keith Shaw: Wow.

Solopreneur—did you coin that?

Michael Todasco: I’m sure I heard it somewhere else.

There’s a Google tool—Ngram Viewer—that shows when words were used historically. I use it when writing period pieces.

You can see if a word existed in the 1930s or not—it’s really helpful.

Keith Shaw: I’m writing that one down.

Michael Todasco: Bonus tip!

Keith Shaw: Got any other projects you’re working on—something fascinating or terrifying?

Michael Todasco: Honestly, cheesecake has consumed my brain.

But seriously, that’s one thing LLMs still don’t get—they haven’t lived.

Like when my daughter was nine, she’d eat pizza by tunneling through the middle—sauce everywhere.

Adults learn to eat around the sides. AI hasn’t eaten a burrito. The internet doesn’t teach that nuance.

If I started an AI company, I’d record human behaviors at an ice cream shop, then sell that training data.

That’s the missing link—actual human behavior, not internet performance.

Even experts like Yann LeCun and Fei-Fei Li are moving on from LLMs to world models.

Keith Shaw: Is that why Meta’s making those glasses—to record real-world data from users?

Michael Todasco: I don’t think it’s the primary reason—but it’s definitely a secondary benefit.

Keith Shaw: I’m not wearing a tinfoil hat... yet.

Michael Todasco: No, but if Meta offered you $200/month to wear those glasses, plus upgrade your internet—you’d probably say yes.

Imagine doing that for 10,000 people worldwide. That’s real data. That’s how we train better AI.

Keith Shaw: That’s my new solopreneur job—record my life, send in the footage.

Michael Todasco: That will be a real job. "Just live life, and wear these glasses." The new TaskRabbit.

Keith Shaw: All right, is there a clear leader in creative AI right now? Or should people still try different models?

Michael Todasco: It depends. If you’re coding and not using Cursor, Gemini 2.5 is very strong.

We did an in-class coding challenge—students using Gemini outperformed others.

For writing, I prefer Claude Sonnet. But GPT-4o is now quite strong at analyzing and improving drafts.

Honestly, for most users—it doesn’t matter. Use what your company offers.

These tools are all really good. The differences are at the margins.

Keith Shaw: Mike, thanks again for joining us. We’ve got to get working on that Golden Girls/ALF comic book.

And don’t forget—you’ve got cheesecake research to conduct.

Michael Todasco: I’m spending the rest of my day obsessed with cheesecake.

Keith Shaw: We’ll report back.

Michael Todasco: Sounds good. Always a pleasure.

Keith Shaw: That’s going to do it for this week’s episode. Be sure to like the video, subscribe to the channel, and leave your comments below. Join us every week for new episodes of Today in Tech. I’m Keith Shaw, thanks for watching.